Research on PCB Defect Detection Based on YOLOv8n-DDS

To address the challenges of insufficient detection accuracy, real-time performance, and lightweight design caused by the diversity of PCB defect features, varying shapes, and sizes, this study proposes an improved algorithm based on YOLOv8n, named YOLOv8n-DDS. First, the Deformable Convolutional Network v2 (DCNv2) module is integrated into the first and second C2F modules of the backbone network. This module dynamically adjusts sampling positions and weights, enabling the model to better capture geometric deformations and complex features of targets. Second, the Dilated Widthwise Residual (DWR) module is introduced into the third and fourth C2F modules of the backbone network. This module employs dilated convolutions to expand the receptive field without increasing parameter count or computational complexity. Finally, the SEAM attention mechanism is incorporated into the detection head (Head) to enhance the model’s focus on critical features and improve recognition capability. Experimental results on a PCB defect dataset demonstrate that the YOLOv8n-DDS model achieves improvements of $26.3 \%, 12.4 \%, 19.4 \%$, and $27.6 \%$ in precision (P), recall (R), mean average precision (mAP) and frames per second (FPS), respectively, while reducing GFLOPs and model size by $13.6 \%$ and $7.7 \%$. compared to the baseline YOLOv8n. These results validate the superior comprehensive performance of the proposed model.